GEC-Agent: Tool-Augmented Large Language Models for Grammatical Error Correction
Abstract: In the era of large language models (LLMs), utilizing these models to address a variety of Natural Language Processing (NLP) tasks has emerged as a focal point of research. However, applying LLMs to the Grammatical Error Correction (GEC) task remains challenging. In this paper, we introduce GEC-Agent, a novel framework designed to effectively leverage the inferential and syntactic capabilities of LLMs while integrating external tools and rule-based approaches to enhance correction accuracy. The framework incorporates grammar and retrieval tools to identify and correct grammatical errors effectively, and implements a reflection mechanism to mitigate overcorrection. GEC-Agent dynamically selects appropriate tools to optimize the correction process and ensures consistency with the original text's style. Our experiments on the CoNLL-2014 and JLFEG datasets demonstrate that GEC-Agent outperforms the few-shot method, using the same large language model, and achieves a higher recall rate compared to existing traditional methods with supervised learning.
Paper Type: Long
Research Area: NLP Applications
Research Area Keywords: grammar error correction, large language model, agent
Contribution Types: NLP engineering experiment
Languages Studied: English
Submission Number: 1938
Loading